Guillaume Lagrange

Yeah there is no harm in that 😄 I am just a little puzzled hahah

There is no official implementation at this time 🙂 LoRA is a simple re-parametrization of the linear layer weights, so this should be possible to implement. We'd like to have...

Hey @wcshds 👋 I was going to update the cubecl dep to the latest to include some fixes, but some wgpu tests are failing with your [merged PR](https://github.com/tracel-ai/cubecl/pull/53). ``` failures:...

Weird 😅 Not sure if the fix is the proper way to address this or if it's just a patch for a more specific issue. It doesn't seem to happen...

Gather op actually translates to `tensor.select(dim, indices)`, which is limited to 1d indices. So the dim inference computes the correct output shape, but the codegen is not implemented for rank...

> I just noticed that burn via crates is only up to version 0.13.2, but the main branch on github is at 0.14. I just tried building the first model...

@alteredoxide as you noted in your first point, `tensor.gather(dim, indices)` is not the equivalent of the `Gather` ONNX op. The closest matches that we have for these ops are: -...

> @laggui Thanks, I clearly pulled on the wrong thread, and falsely assumed your previous statement about mapping to `select` was more figurative than literal (as though the logic was...

Yes we generated the onnx file for the tests via this python script. There are also 2 other scripts used to generate different configurations of the gather node, so you...

> * **I added a `gather_onnx` method to the `tensor/api/numeric.rs` module** -- this is the situation I had in mind regarding the issue of naming: that module already has `gather`...